Learning-Based Named Entity Recognition for Morphologically-Rich, Resource-Scarce Languages

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 354–362, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Learning-Based Named Entity Recognition for Morphologically-Rich, Resource-Scarce Languages Kazi Saidul Hasan and Md. Altaf ur Rahman and Vincent Ng Human Language Technology Research Institute University of Texas at Dallas Richardson, TX 75083-0688 {saidul,altaf,vince}@hlt.utdallas.edu Abstract Named entity recognition for morpholog- ically rich, case-insensitive languages, in- cluding the majority of semitic languages, Iranian languages, and Indian languages, is inherently more difficult than its English counterpart. Worse still, progress on ma- chine learning approaches to named entity recognition for many of these languages is currently hampered by the scarcity of annotated data and the lack of an accu- rate part-of-speech tagger. While it is possible to rely on manually-constructed gazetteers to combat data scarcity, this gazetteer-centric approach has the poten- tial weakness of creating irreproducible results, since these name lists are not publicly available in general. Motivated in part by this concern, we present a learning-based named entity recognizer that does not rely on manually-constructed gazetteers, using Bengali as our represen- tative resource-scarce, morphologically- rich language. Our recognizer achieves a relative improvement of 7.5% in F- measure over a baseline recognizer. Im- provements arise from (1) using in- duced affixes, (2) extracting information from online lexical databases, and (3) jointly modeling part-of-speech tagging and named entity recognition. 1 Introduction While research in natural language processing has gained a lot of momentum in the past several decades, much of this research effort has been fo- cusing on only a handful of politically-important languages such as English, Chinese, and Arabic. On the other hand, being the fifth most spoken lan- guage1 with more than 200 million native speakers residing mostly in Bangladesh and the Indian state of West Bengal, Bengali has far less electronic resources than the aforementioned languages. In fact, a major obstacle to the automatic processing of Bengali is the scarcity of annotated corpora. One potential solution to the problem of data scarcity is to hand-annotate a small amount of data with the desired linguistic information and then develop bootstrapping algorithms for com- bining this small amount of labeled data with a large amount of unlabeled data. In fact, co- training (Blum and Mitchell, 1998) has been suc- cessfully applied to English named entity recog- nition (NER) (Collins & Singer [henceforth C&S] (1999)). In C&S’s approach, consecutive words tagged as proper nouns are first identified as poten- tial NEs, and each such NE is then labeled by com- bining the outputs of two co-trained classifiers. Unfortunately, there are practical difficulties in ap- plying this technique to Bengali NER. First, one of C&S’s co-trained classifiers uses features based on capitalization, but Bengali is case-insensitive. Second, C&S identify potential NEs based on proper nouns, but unlike English, (1) proper noun identification for Bengali is non-trivial, due to the lack of capitalization; and (2) there does not ex- ist an accurate Bengali part-of-speech (POS) tag- ger for providing such …

Content truncated to 500 words